Core Concepts
A novel framework, AdaIR, that leverages adapters to efficiently adapt a shareable foundation model to diverse image restoration tasks, achieving comparable performance with significantly fewer parameters and reduced training time.
Abstract
The paper proposes AdaIR, a framework that integrates adapter modules into a shareable foundation model to enable efficient adaptation to various image restoration tasks. The key insights are:
The framework consists of a pre-training phase to uncover shareable components across restoration tasks, followed by a fine-tuning phase where only lightweight, task-specific adapter modules are trained.
The authors analyze the influence of different pre-training schemes on the performance of downstream tasks, providing valuable insights for future advancements.
Extensive experiments demonstrate that AdaIR achieves favorable performance on multiple restoration tasks, including denoising, deblurring, deraining, and super-resolution, while utilizing significantly fewer parameters (1.9 MB) and less training time (7 hours) compared to existing methods.
The paper highlights the benefits of parameter-efficient tuning techniques, such as adapters, for the low-level vision domain, which has been relatively unexplored compared to high-level vision tasks.
Stats
AdaIR requires only 1.9 MB of tunable parameters, which is less than 8% of the parameters of the Restormer baseline.
AdaIR can be fine-tuned on each restoration task in just 7 hours, significantly faster than training Restormer and PromptIR from scratch.
Quotes
"AdaIR tunes lightweight adapter modules Ar to learn task-specific knowledge for restoring Ir^HQ by eliminating the artifacts present in the Ir^LQ."
"Extensive experimental results show that AdaIR achieve outstanding results on multi-task restoration while utilizing significantly fewer parameters (1.9 MB) and less training time (7 hours) for each restoration task."